Self-correcting retrieval
Document relevance, answer grounding, and usefulness gates control the generation loop; failed checks trigger corrective regeneration, query rewriting, or web fallback.
COST-EFFICIENT AI ENGINEERING
Reliable AI systems built for measurable outcomes.
Agentic RAG, Enterprise Agent Architecture, and Multi-Agent Workflow Automation — designed around quality gates, cost control, privacy boundaries, observability, and real business impact.
Agentic RAGLLM EvaluationModel Routing
Enterprise AgentsPrivacyMulti-Agent Automation
engineering profile
focus Agentic AI / Applied AI
optimize quality × cost × reliability
build RAG · agents · workflows
validate evals · tests · observability
protect privacy · budgets · boundaries
84.5%
lower measured LLM inference cost
Task-level model routing vs. legacy single-model baseline
100%
benchmark pass rate
Improved from 98.3% to 100% (180/180)
~50%
hardware cost reduction
Commercial workstation redesign
24
behavioral evaluation cases
Across 6 RAG scenarios
01
AI SYSTEMS · COST · EVAL · SECURITY
Enterprise Agentic RAG
A self-correcting CRAG-style document Q&A system with conditional routing, quality gates, bounded retries, web fallback, privacy controls, and fully local execution.
Model routing & benchmark harness
Task-level model routing cut measured inference cost by 84.5% while improving benchmark pass rate from 98.3% to 100% (180/180). A protocol-driven benchmark harness adds frozen inputs, fresh-process isolation, observation-level model identity / usage / cost audits, and quality-first decision gates.
Privacy & reliability boundaries
Per-run LLM/web budgets, secret redaction, prompt-injection hardening, metadata-only observability, graceful degradation, Privacy Mode, and Fully Local Ollama mode.
Route→Retrieve→Grade→Generate→Ground→Verify
180/180optimized benchmark score
6 taskstask-level model allocation
3 gatesrelevance · grounding · usefulness
LocalLLM + embeddings + retrieval via Ollama
02
ENTERPRISE AGENT ARCHITECTURE
Enterprise Office Agent
A deterministic-first agent layer that routes free-text requests across seven enterprise workflows, using LLMs only where language synthesis adds value.
Knowledge Q&AEmail SummaryCalendar Lookup
Ticket AssistantDaily BriefingMeeting PrepWorkflow Approval
Deterministic-first routing
Core office workflows remain reproducible and offline-capable by default; Knowledge Q&A delegates to the RAG engine through a thin adapter.
Controlled LLM action surface
Email Summary and Daily Briefing can use bounded LLM assists, but the model cannot send email, approve workflows, or mutate business data.
Observable & testable
The workspace exposes execution paths, timings, run settings, counters, and privacy state. Validation recorded 350 Office Agent tests and 1,103 repository-wide tests.
03
MULTI-AGENT AUTOMATION · SAAS INTEGRATION
Business Development & RevOps Automation
A Manager–Worker workflow that turns requirements documents into prospect research, CRM records, and personalized outreach drafts while preserving human review before sending.
Requirements PDF→Manager Agent→Prospecting
→RevOps+SDR→CRM + Gmail Drafts
Manager–Worker orchestration
A Business Development Manager parses ICP / requirements PDFs, invokes Prospecting once, then fans out RevOps and SDR sub-agents independently per prospect.
Real SaaS integrations
Firecrawl and Hunter support company/contact research and work-email discovery; Pipedrive creates Organizations, Persons, and Leads; Gmail creates personalized drafts.
Explicit data contracts
Structured Output, subworkflow input schemas, conditional branching, and isolated side-effecting steps separate LLM reasoning from CRM writes and draft creation.
n8nOpenAIFirecrawlHunterPipedriveGoogle DriveGmail
04
COMMERCIAL ENGINEERING EXPERIENCE
Full-Stack Engineering · Dental SaaS
Full-Stack SaaS Development
Sole full-stack engineer for a dental SaaS application, responsible for product design, system architecture, frontend and backend development, debugging, maintenance, and iterative feature delivery.
Cost-focused redesign
Redesigned a mobile dental-scanning workstation to improve portability and simplify the system architecture, reducing related hardware costs by approximately 50%.
Business impact
Delivered tens of thousands of dollars in confirmed savings during the initial rollout, with additional six-figure projected savings at full scale.
05
AI-NATIVE ENGINEERING WORKFLOW
Architecture-first, AI-assisted development
Coding agents are used for rapid codebase onboarding, architecture mapping, implementation, review, and validation — with task-based model selection and cost-aware workflows.
01Scan codebase & build architecture map
02Identify critical execution paths
03Write spec & constrain implementation
04Review diffs, tests, and system boundaries
05Use stronger models only where complexity justifies cost
06
TECHNICAL STACK
Tools used to ship and evaluate AI systems.
AI / LLM
LangGraph · LangChain · RAG / CRAG · LLM Evaluation · Model Routing · OpenAI · Together · Ollama · Chroma · Tavily · n8n · MCP
Languages
Python · JavaScript · SQL
Full Stack
React · Node.js · Koa · MySQL
Engineering
Git · PowerShell · uv · pytest · CI/CD · Codex · Claude Code